Jiaying Lu, Carl Yang.

Taxonomies and knowledge graphs (KGs), which represent real-world entities’ abstract concepts and properties/behaviors/facts, constitute the essential information in knowledge bases (KBs). However, most existing KBs are constructed under the closed-world assump- tion, which often corresponds to a fixed schema and requires ad-hoc canonicalization to integrate new knowledge. To empower KBs towards easy accommodation of emerging entities and relations, we propose to create open-world TaxoKGs based on existing au- tomatically constructed taxonomies and open KGs, where taxonomies serve to provide a loosely-defined schema and mitigate the reliance on ad-hoc canonicalization. To further improve the completeness of TaxoKG, we collect several new benchmark datasets towards the development of HakeGCN, an innovative hierarchy-aware graph-friendly model for TaxoKG completion. Through extensive experiments, we demonstrate HakeGCN to outperform various state-of-the-art KB completion methods on both taxonomy concept prediction and KG relation prediction tasks based on both standard metrics and human evaluations. The benchmark datasets and the implementation of HakeGCN are available at https://github.com/lujiaying/Open-World-TaxoKG-CoLearning.

Citation

@inproceedings{
lu2022open,
title={Open-World Taxonomy and Knowledge Graph Co-Learning},
author={Jiaying Lu and Carl Yang},
booktitle={4th Conference on Automated Knowledge Base Construction},
year={2022}
}